Capture Notes

Preprints.org DOI:10.20944/preprints202602.1188.v1. The page states this version is not peer reviewed.

Why Collected

Highly relevant because it directly models cross-agent propagation of poisoned memory in multi-agent collaborative environments.

Key Metadata

• Submitted: 2026-02-13
• Posted: 2026-02-14
• Type: Preprint, not peer reviewed

Collection Summary

The paper proposes a memory poisoning detection and repair method for MAS using source credibility, semantic consistency, evidence graphs, contrastive learning, isolation, rewriting, and conflict resolution. It states that experiments used 60 collaborative tasks, about 210,000 memory records, and 12,000 injected poisoned samples, reporting lower misbehavior and reduced cross-agent propagation.

Security Relevance

• Direct candidate evidence for cross-agent propagation, directed memory propagation graph, and propagation-path intervention.
• Strongly overlaps with the metrics and defense variables in [[04_Research_Questions/RQ - MAS Misevolution Propagation Control]].
• Because it is a non-peer-reviewed preprint and some claims are unusually specific, ingest should preserve uncertainty.

Suggested Ingest Focus

• Create evidence around memory propagation graph and repair loop if source is accepted as usable.
• Mark trust tier as preprint and confidence as medium-low until peer-reviewed version or code/data are verified.
• Compare with [[raw/papers/from-spark-to-fire-error-cascades-2026]] and [[raw/papers/memory-poisoning-secure-mas-2026]].